A computerized method for retrieving documents from a text corpus in response to a user-supplied natural language input string, e.g., a question. An input string is accepted and analyzed to detect phrases therein. A series of queries based on the detected phrases is automatically constructed through a sequence of successive broadening and narrowing operations designed to generate an optimal query or queries. The queries of the series are executed to retrieve documents, which are then ranked and made available for output to the user, a storage device, or further processing. In another aspect the method is implemented in the context of a larger two-phase method, of which the first phase comprises the method of the invention and the second phase of the method comprises answer extraction.